Modeling Online Ideological Community Dynamics with Recurrent Variational Graph Auto-Encoders

Social media platforms have become powerful channels for shaping and influencing public opinion. The rapid dissemination of information online, combined with the everchanging nature of social networks, fosters dynamic patterns of polarization and radicalization that profoundly impact collective behaviors. These platforms offer a rich, large-scale, real-time view of public beliefs, capturing interpersonal interactions and responses to external events. While progress has been made in belief representation learning on static social graphs and temporal representation learning for generic graphs, the former overlooks the dynamic, temporal nature of belief evolution, and the latter fails to account for the unique cues present in temporal social interaction graphs. To bridge this gap, we propose a novel method (RIVGAE) for dynamic belief representation learning that explicitly tracks the evolution of beliefs within time-evolving social graphs, thereby enabling social applications, such as the detection of polarization. Our method integrates recurrent neural architectures with graph-based belief representation learning, capturing the complex interplay between changes in the network and shifts in ideologies of a group of people. Our approach shows significant improvements over generic temporal graph learning methods in ideology classification and interaction prediction through empirical evaluations on diverse social media datasets. It effectively reveals key phenomena, such as ideological divides and processes like polarization and the emergence of radical clusters.